Event-driven implementation of deep spiking convolutional neural networks for supervised classification using the...

Area of Science:

• Computer Science
• Neuroscience
• Artificial Intelligence

Background:

• Moore's Law has driven neural network advances, but silicon limits necessitate power-efficient computing paradigms.
• Neuromorphic hardware, inspired by the brain, offers a solution for energy-efficient learning, particularly for spatio-temporal data.
• Spiking neural networks (SNNs) show promise in matching classical neural network performance on supervised tasks.

Purpose of the Study:

• To demonstrate the feasibility of implementing state-of-the-art neural network models on neuromorphic hardware for digit recognition.
• To evaluate two approaches: direct conversion of classical networks and training SNNs from scratch.
• To introduce a novel method for reducing spike rates in SNNs without compromising accuracy.

Main Methods:

• Implementation and simulation of neural network models on the SpiNNaker 103 neuromorphic platform.
• Utilizing both the MNIST and event-based NMNIST datasets for digit recognition tasks.
• Comparing direct conversion of artificial neural networks (ANNs) to SNNs against training SNNs natively.

Main Results:

• Achieved digit recognition performance comparable to state-of-the-art results on both datasets.
• Successfully implemented and tested models on neuromorphic hardware.
• Developed a method reducing spike rates by up to 34x for fully connected architectures, preserving accuracy.

Conclusions:

• Neuromorphic hardware is capable of running sophisticated SNN models for complex tasks like digit recognition.
• The proposed spike reduction method enhances the power efficiency of SNNs on neuromorphic platforms.
• SNNs demonstrate significant potential for efficient spatio-temporal information encoding in embedded applications.